Prévue : Luan Anh NGUYEN-TRAN

Jury

Abstract

Generative artificial intelligence is increasingly embedded in higher education as an AI-based information system whose outputs can enter students’ learning tasks. Through prompt-based interaction, generative AI chatbots produce textual responses that students may use to explain concepts, reformulate content, prepare assignment answers, organise ideas, translate materials, or check their understanding. These outputs are often useful and convincing, yet uncertain: they may be incomplete, inaccurate, generic, poorly aligned with course expectations, or difficult to verify. This creates a central problem for Information Systems research: how users rely on the outputs of a complex and partly opaque AI system after adoption, when those outputs become part of situated work practices. This dissertation examines this problem through the evolution of students’ trust in generative AI during learning practice, focusing on how students perceive its role, regulate their trust, and reshape boundaries between AI assistance, student work, teacher validation, and pedagogically legitimate knowledge. The empirical study adopts an interpretive qualitative approach and is situated in a Microbiology course at a Vietnamese public university, following 61 second-year Food Technology students across three assignments during one semester. Students were not required to use generative AI, but declared their use and provided prompt documentation when they did. The empirical material combines assignment submissions, AI-generated artefacts, responses to an open-ended questionnaire, one teacher interview, assignment grades used as contextual indicators, and three student case vignettes constructed from multi-source materials. The findings indicate that students perceived generative AI as both capable and fallible: useful for explanation, translation, organisation, reformulation, and assignment preparation, yet sometimes wrong, incomplete, generic, or poorly aligned with the course. Trust therefore evolved through situated reliance, as students copied, modified, verified, selected, combined, or rejected AI-generated outputs. Three forms of reliance were observed: stable direct reliance, partial reliance, and variable heavy reliance. Partial reliance was dominant, indicating that generative AI had entered the sociotechnical organisation of learning without simply replacing course materials, teacher validation, or student judgement. From a boundary theory perspective, trust shaped how far AI-generated outputs could move from external assistance into assignment preparation, submitted work, and pedagogically legitimate knowledge. The dissertation contributes to Information Systems research by conceptualising trust evolution in generative AI as situated reliance. It explains how recognised fallibility reorganises reliance through verification, modification, selection, and limitation rather than interrupting use. It also contributes to boundary theory by conceptualising trust as a mechanism through which AI-generated outputs move between assistance, student work, teacher validation, and pedagogically legitimate knowledge.